Blockchain-Empowered Vehicular Intelligence: A Perspective of Asynchronous Federated Learning

Blockchain-empowered federated learning is a promising learning framework, which mitigates several potential security threats in learning. However, in the Internet of Vehicles, the asynchronous network puts higher requirements on blockchains. Specifically, due to the asynchronous transaction updates, traditional consensus mechanisms require nodes to frequently coordinate to reach a consensus on the global order of transactions. This strong consistency brings excessive computing time and low efficiency to federated learning. Existing solutions completely relax the consistency of transactions, which, however, reduces the persistence and traceability. Therefore, we propose a lightweight permissioned blockchain with partial consensus, which reduce the coordination among nodes to reduce the system overhead. First, we run the consensus of transactions for global models and relax the strong consistency for local models, which are stored in parallel in real-time without coordination among nodes. Accordingly, we provide relative persistence to ensure the traceability of local models. Then, due to the orderless transactions, we use smart contracts, instead of time stamps, to control the staleness weight of local models in aggregation to reduce the vulnerability. Experimental results show that our scheme effectively improves the performance of blockchain-empowered systems and overcomes the challenges of asynchrony to the security of vehicular networks.